The Rise of the Full-Stack Builder and Hyper-Leveraged Generalist with Microsoft CEO Satya Nadella
In a Nutshell
Satya Nadella argues that AI success now requires delivering tangible value through ecosystem platforms rather than single models, with companies building private evaluations, multi-model harnesses, and company-specific agents on top of foundation models. The biggest opportunity lies in reconceptualizing work as "meta work"—where teams build agentic systems to manage operations like Azure networking, and generalists gain massive leverage by creating applications rather than documents. Long-term, Microsoft positions itself as the operating system layer for intelligence, enabling every organization to compound human and token capital while ensuring communities see real economic benefits from the infrastructure buildout.
These notes were generated by AI and may contain inaccuracies.
The world will be skeptical of tech companies claiming to have the future figured out. Companies must deliver tangible benefits because AI now represents too much of the economy for vague promises. True ambition means making the impossible possible.
Nadella draws inspiration from the Azure networking team, noting that Microsoft built more Azure capacity in the last 15 months than in the first 15 years combined. The goal is not to manually manage Azure networking but to build agentic systems that handle it autonomously.
The methods for accessing information, educating oneself, and staying continuously updated have fundamentally changed. The next major startup opportunity could involve building a new university or novel pedagogy that helps people move through curriculum and achieve economic opportunity.
Nadella's biggest reflection from Build is to conceptualize AI as an ecosystem play rather than focusing on a single model or single platform. Having experienced four major platform shifts at Microsoft, he defines a platform by its ability to create more value outside the platform than what is captured within it.
The keynote demonstrated how any company, whether AI-native or traditional enterprise, can participate as a first-class participant by pointing to AI they created themselves. While companies will use other people's AI, the critical question is providing the path, recipe, stack, tooling, and clarity on what is valuable.
Microsoft's MAI models were built with a focus on clean lineage, beginning with pre-training using high-quality data and conducting ablations to remove problematic content. The challenge of building clean lineage models has increased due to the volume of data that needs filtering.
Many open-weight models perform well on one or two benchmarks but fail in practice. The MAI models are designed so that even a small 5B parameter model can effectively hill climb when given the proper scaffold.
The key objective is finding the cognitive core. Starting with clean lineage enables companies to use models not just as generalists but to create their own specialists by building hill-climbing scaffolds around them.
The approach involves providing not just the model but a hill-climbing scaffold around it. Companies then build reinforcement learning environments, collect traces, and most importantly maintain private evaluations.
Public evaluations can be maxed out and are no longer critical differentiators. Each company needs its own private eval that truly measures what matters to them.
The end-to-end platform story around models includes training strategy, scaffolding, trace collection, and private evaluation infrastructure.
A new frontier concept involves adding temporality to model usage. The Lando Lakes demo demonstrated this by using GPT-5.5 to collect traces, then using a 5B reasoning model to achieve higher performance than the original larger model.
This represents another dimension of operating at the frontier beyond simply using the largest available models.
Nadella reflects on what he wishes he had understood two to three years ago when the OpenAI partnership began and when the scaling laws paper influenced strategy. The core insight was that intelligence scales logarithmically with compute, and throwing more compute at transformers produces capability gains.
What was underestimated was the real-world complexity of deploying these systems to actually deliver value. Benchmarks measure outcomes but the true evaluation occurs when people in the real world accomplish unique things they value in measurable ways.
The industry has been overly focused on token maximization as an artifact of not thinking about tokens as a means to create value at every step.
Coding has demonstrated large-scale impact, to the point that IDEs now need rebuilding. Microsoft launched tooling supporting 100 agent sessions, creating excessive cognitive load that requires new UI paradigms. The chat-only artifact model is insufficient, necessitating canvas-based interfaces even for code work.
Beyond coding, value is emerging in augmenting glue work and judgment tasks. Tools like AutoPilot and Copilot demonstrate how agents that are long-running and durable can scale human capital work similarly to how coding agents scaled development.
In the near future, agents will work overnight on delegated authority, requiring new audit mechanisms for humans to review what agents accomplished.
A harness defines the models, data, and tools with loops across all three. Each Microsoft product, whether GitHub Copilot, security Copilot, M-dash, or scientific discovery tools, functions as a multi-model harness with tool access for progressive disclosure that maintains token efficiency.
Rich context preparation is critical. The amount of work required to prepare the context layer so plans execute efficiently represents where the magic occurs.
GitHub harness is being used across all Microsoft products and is available in Foundry. Companies can use their own Llama harness or any open harness, training with their tools, multiple models, and context.
The value proposition is that training harness plus tools plus model together produces evaluation gains. The M-dash launch provided existence proof that a multi-model harness can outperform Mythos in finding bugs and vulnerabilities not detected by other systems.
This challenges the premise of independent frontier labs that assume first-party products drive majority revenue while enterprises primarily consume via APIs. Microsoft operates with both first-party products and enablement products.
Platform builders can have successful first-party products, as demonstrated historically with Windows, SaaS, and cloud. The critical requirement is that this should not limit others from achieving similar success.
Network effects around intelligence differ because models learn from few samples rather than requiring massive datasets. This shifts the game toward protection mechanisms.
Every company having private evaluations may become the biggest form of IP. A private eval enables using a frontier model to hill climb without leaking traces.
An asset test determines control: if a company has a private eval and can switch from model A to model B while maintaining performance gains, they are in control. If they cannot switch models successfully, they lack control.
Harness decisions become critical because open harnesses that accept all models, combined with private evals, context, and tools, enable hill climbing. These represent the skills AI-native startups, SaaS companies, and enterprises all need.
Microsoft's evolution from operating systems company to cloud company may have a third act as the harness and evaluation layer that enables every company to have frontier intelligence with their own data.
The mission is enabling everyone to operate at the frontier with their frontier intelligence. This creates stable equilibrium where companies achieve terminal value by continuously compounding on a platform that improves over time.
Historical parallels include Windows enabling Adobe and Autodesk, and Jensen building CUDA on top of DX. The platform layer allows others to extend and build their own intelligence layer.
Every company will have both human capital and token capital. The question is how to compound the two. Human ability to find gaps remains essential for value creation.
This represents expressing new forms of human agency and ambition even as token capital increases. In Teams, agents and humans working together generate traces that become valuable context for how the enterprise creates value.
This enables training company-specific veteran agents rather than generalist models. The tacit knowledge captured through traces may eventually appear on balance sheets in ways human capital never could.
Traditional SaaS captured workflow through data models, business logic, and UI layered on top. Now the vertical stacking can be relitigated.
Data models underneath SaaS applications remain valuable and should not be reinvented. General ledgers should remain general ledgers without new schema creation. Entity relationships are robust and worth preserving.
Business logic embedded in tools like PowerBI semantic models represents work worth preserving and making available.
The challenge for SaaS business models is learning to unbundle existing packages and rebundle in new ways while discovering new business models.
Microsoft 365 with Work IQ exposes what may be the most important database in a company that was never used as a database because it was captive to applications. Email, Teams, Word, Excel, PowerPoint, and SharePoint all operated on this data.
A demonstrated use case involves going to a GitHub repo and asking what changes should be made based on design meetings attended the previous week. The system reviews transcripts and returns a plan to modify the codebase.
This represents value creation opportunities 10x larger than previous paradigms, but requires rearchitecting serving infrastructure since mailbox-serving systems cannot serve agents.
Per-user pricing emerged as an artifact of budget certainty needs. Per-user represents a set of entitlements to usage.
The first bundling approach takes usage and bundles it into per-user stacks sold as subscriptions. Subscriptions and per-user pricing will persist.
Consumption-based pricing represents the next major shift. Outcome-based pricing is theoretically appealing but faces resistance once outcomes materialize, as companies resist sharing royalties.
GitHub Copilot illustrates the adjustment challenge: originally priced per user for interactive code completion, it now supports 10,000 agents running continuously, requiring consumption metering alongside per-user options.
One full budget cycle is needed to observe equilibrium emergence. The decision framework is whether the marginal cost of building and maintaining something internally exceeds the cost of acquiring it.
Maintenance considerations matter because AI-discovered security issues require fixes that burn tokens. Responsibility assignment becomes a cycle requiring careful analysis.
The initial excitement about generating software will give way to questions about what software should be generated internally versus acquired, and how to compose both into agentic workflows with retained agency.
Vendors demonstrating inflexibility will face low tolerance. Vendors showing flexibility and delivering value will continue selling software under different business models.
Building software has made it possible for even the CEO of a company like Microsoft to build. GitHub Copilot and the new sessions app have made it possible to have agency over artifacts that previously felt untouchable. Nadella recalls joining Microsoft and needing to examine codebases like Cutler's to learn good C++ code. The ability to be more full stack up and down is now dramatically improved.
Nadella has been building long-running foundry agents and autopilots. He recently built an agent that continuously monitors his own chief of staff autopilot, which will also exist in Scout. He took Work IQ, asked it to build a foundry long-running agent, stored all the memory using Rayfin as a backend as a service, and it built the agent. He was able to publish it to Teams. The ability to complete an end-to-end project like this is described as pretty miraculous.
The question was posed whether in four or five years there will be essentially four engineering roles: people managing agents, forward deployed engineers or FTEEs, security engineers, and people working on large-scale infrastructure for a small number of services, with everything else collapsing into the agentic world.
Nadella noted that LinkedIn structurally changed and built up a new discipline called full stack builder, bringing together people from design, product management, and front-end engineering. The design person still has the design edge and the front-end person still has the front-end edge, but they can give themselves bigger scope in roles rather than being confined to one role.
Infrastructure has become very critical. Even for the Excel team, building the RLE in which a reward can be learned is one of the hardest infrastructure problems. New talent is needed, including distributed systems people even in what was considered an end-user app team.
The world will always have specialists, but the generalist role is going to be the most exciting because the leverage of a generalist is where maximum returns will be seen. Knowledge work that previously involved creating a Word document or spreadsheet can now involve building an app. Generalist skills have gotten higher leverage.
This is music to the ears of CEOs and VCs. It is a golden age for idea people with a lot of agency.
Nadella's partner Mike Renol, who started his career at Microsoft, wrote an essay noting that this is an age where organizations can be much more ambitious given the pace of the environment and how quickly users and companies are open to adopting new technologies.
Nadella referenced Kevin Scott's line that when you can make hard things easier, that is one point of leverage, but true ambition is about making the impossible possible. What is missing in organizations is a conceptual model of what can be built that was previously impossible.
The Azure networking team came to the realization that their job scaling Azure capacity, which in the last 15 months exceeded what was built in the first 15 years, would not work without reconceptualizing their work. They determined their job was not to do Azure networking but to build the agentic system that does Azure networking. They manage over 500 fiber operators and the WAN. Fiber operations is a physical operation involving repairs and responses to issues. They built an agentic system with a character called Miles that handles this work. They started screaming for more tokens rather than headcount, saying they needed tokens to manage their operation. They took their work and made it meta, and that meta work is now their new work.
In the 80s, if someone had said four billion people would get up and start typing, the model would have been needing four billion typists. Instead, people are doing knowledge work. The key is giving organizations permission to do new types of meta cognition and meta work using new tools to change the outputs that matter and make the impossible possible. Completing the connective tissue across these areas is where a lot of enterprise value will get created.
The data center buildout is extraordinary and nothing like this has happened before. At this point it is clear that unless the industry is very principled about ensuring the benefits are felt in real ways at the community level, there will be problems. This is not just a campaign. People need to see that this is not changing energy prices for them, and in fact long-term there will be a better grid and prices may come down. Water consumption involves closed loop systems where water is being replenished.
Companies must invest in training, jobs, and tax base. The amount of jobs created during and after construction and the tax base in the community are the least talked about aspects. If this is real, there will be permission. If not, there will not be permission. It is as simple as that.
Communities should be skeptical and ask hard questions, and the industry needs to do the hard work to earn that permission. In human history, using a lot of energy but creating a lot of value for society has been a fantastic story. If that does not happen, it has not been great.
Nadella is a firm believer that if there is a token economy that drives productivity, economic growth, broad participation, and better health outcomes, things will be in a great place.
ROI might be easier to see in communities first before in enterprise. Both sides come back together. People in communities are going to be employed and participants in the real economy. If the broad economy is doing well and communities are doing well, the dots get connected through market forces. Evidence must be visible. It cannot be about any one company, but must be about broad economic growth and community permission.
The most critical thing is telling the story and making it real that everybody has a real shot to participate as a first-class participant in this new economy. In the next 12 to 18 months, there needs to be a way for people to say they get it. There is going to be tremendous capability and infrastructure, but people need to see what is going to happen, whether in health outcomes, ability to create a startup, or ability to run a local store more efficiently.
The world is going to be very skeptical of tech and tech companies that say trust us and the future will be glorious. Tangible benefits must be delivered. Politicians winning elections because they have advocated for that will be a key adjustment.
A simple framework for the broad benefit of AI beyond communities working in technology includes wealth creation happening in startups and large companies, healthcare with companies like Open Evidence, and education, which seems like an obvious good where as much impact as expected has not yet been seen.
Nadella met with the founders of Alpha School and learned about rethinking what education really looks like. It is important, and traditionally done things are not necessarily less important. A Stanford class on Asian guidelines for CS was referenced, where people still need to learn concepts like applying softmax appropriately rather than just fixing a training run. Learning concepts is important and will be critical.
However, the way incentives are created, what credentials are valued, and what employment opportunities exist for those credentials need complete change given how access to information, the ability to educate oneself, and the ability to continuously update have changed so much.
The next big startup and success story could be someone who builds a new university or a new pedagogy of how to get someone through a curriculum and find economic opportunity. That is highly valuable.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
